基於機器視覺與深度學習的面板切割缺陷檢測

Machine Vision and Deep Learning Based Scribe Defect Detection for LCD

林祐誠(You-Cheng Lin) ； 黃俊翔(Chun-Hsiang Huang) ； 羅鈺淳(Yu-Chun Luo) ； 何昭慶(Chao-Ching Ho)

《科儀新知》 235期 (2023/06) Pp. 66-75

摘要

近年來資通訊產品需求增加導致TFT-LCD面板產業的蓬勃發展，但TFT-LCD面板的製程檢測仍然仰賴人工目測檢查，而人為評估瑕疵質量通常是主觀的且不穩定的，因此導入自動光學檢測（automated optical inspection，AOI）取代人工檢測是一項重要的自動化課題，且不同表面特性的材料需對應不同的取像照明方式，若取像照明方法不穩定會影響缺陷檢測過程中的整體穩定性。為此本研究開發一套用於TFT-LCD面板線上的即時自動光學檢測系統，利用影像辨識基板切割後耳料是否脫離正常，避免TFT-LCD面板因切割不良未檢出並流入後段製程，導致面板破裂或後端機台損毀，幫助提升製程良率，減少設備損毀之不良影響，最後加入深度學習來提升瑕疵檢測的效率，結果顯示加入深度學習的影像預測速度效率相比傳統方法提高了38%，並達到100%的預測準確率。

並列摘要

In recent years, the demand for information and communication products has increased, leading to a thriving TFT-LCD panel industry. However, the process of inspecting TFT-LCD panels still heavily relies on manual visual inspection, which can often be subjective and unreliable. Therefore, replacing manual inspection with automated optical inspection (AOI) is an important issue. Since different materials with varying surface characteristics require different illumination methods, unstable illumination can affect the overall stability of the defect detection process. In this study, we developed a real-time automatic optical inspection system for TFT-LCD panels. The system uses images to identify whether the material is abnormal after the TFT-LCD substrate is cut, preventing poorly cut TFT-LCD panels from entering the back-end process, which can result in panel breakage or damage to the back-end machinery. The system helps to improve process yields and reduce equipment damage. The results show that the deep learning image prediction speed is 38% more efficient than the traditional method, and the prediction accuracy reached 100%.